precompute generates invalid code when I give it differing index variable names

[Infinoid]

This works: taco 'A(i) = B(i)*1' -s='precompute(B(i)*1,i,i)'
This breaks: taco 'A(i) = B(i)*1' -s='precompute(B(i)*1,i,j)'

These two commands generate quite different output. The output of the first one appears correct at first glance; the output of the second one has several issues:

• somehow, the size of the workspace stopped being equal to the input size, and is now hardcoded to 42
• the main copy loop now loops over the workspace var, not the input var, and is more complicated as a result
• the output loop now refers to two index variables that are never defined (iA and iworkspace)

If you then add -f=B:s so that the B tensor is sparse, the output of the first case still appears correct, but the second one now has an additional problem:

• the main copy loop now writes to workspace positionj. j is initialized to i, but never updated from one loop iteration to the next

The full outputs are here. This was using a Debug build of taco with no features (openmp, cuda, or python).

There's some overlap between this and comments in issue #347. But I think the isLowerable failure is a separate issue.

[rohany]

I don't know much about this, but it looks like the two commands you have written are the same.

[Infinoid]

The difference is in the precompute() scheduling directive, the ,i or ,j at the end.

Giving it the same IndexVar name twice works. Giving two different names breaks.

[rohany]

Ah my bad, I stared at this for a while and couldn't tell that the i and j were different.

[weiya711]

This is a known bug and happens whenever i and iw differ when calling precompute(expr, i, iw) differ. We are working on fixing this code generation issue now. @Infinoid you are correct that the isLowerable failure from #347 is different since the i and iw differing bug should still generate (albeit incorrect) code without any assertion errors

[Infinoid]

Looks like #459 fixed this. Thanks @nirvikbaruah!